FranciscoCarlos

BTW also saves 10k - 20k tokens per fresh context depending on config

Cooked for 1s · done 11:01 AM Opus 5.5 14k ctx ~0.088%/call 5h 2.0%

blank: custom Opus 5.5 4k ctx ~0.085%/call 5h 2.0%

blank: blank Opus 5.5 1k ctx ~0.084%/call 5h 2.0%

For anyone interested in ditching anthropic's system prompt alltogether

https://github.com/franciscocarloserra/claude-code-context-d...

    Claude Code v2.1.319
  Opus 5.5 with medium effort · Claude Pro

 what's my email                                                                                

● ----@gmail.com

    Claude Code v2.1.319
  Opus 5.5 with medium effort · Claude Pro

 /clear                                                                                         
 /custom 
  ⎿  blank: mode: off

 do you have my email in context?                                                              

● No [...]

Author here, feel free to ask anything

Still a crutch around incomplete pre-training data, a ton of useful knowledge wont be properly represented on the weights.

Also a great potential indicator on when the pretraining run for the model started.

Willing to scale it up if there's interest

It works surprisingly well https://franciscocarloserra.github.io/mp3tomidi/

Ungated quant https://huggingface.co/DamRsn/muscriptor-gguf

Yeah, it can be worked around with in-context instructions but they still tend to scope, delegate and architect without awareness of the fact that agents exist and work. It's an actual big deal IMHO, opus 5.5 is the most updated AFAIK.

They know it and they try to cover it up with a fake reported cutoff date. Claude reports June '26 but it has only partial 2026 data (tested it against the API and it's 25 as well). GPT refuses to answer but it's old 25 data.

Seems like a rushed RL run on top of an old model to jump into the hype train, seems reasonable since they hold a fantastic distribution moat. Even if it's not SOTA a billion people will use it and they'll collect data-points for future iterations. Same as muse.

How is a model that has never seen an agent on it's pretraining data going to reasonably plan and fan-out tasks to multiple agents? Or engineer software for agents to edit semi-automatically, it has zero awareness of it's capabilities in it's internal representation.

They need to pretrain on fresh data.

Gemini 4 Argon

As of my knowledge cutoff, three of the top state-of-the-art large language models are Google's Gemini 2.5 Pro, OpenAI's o3 (and GPT-4.5), and Anthropic's Claude 3.7 Sonnet.

Trivial to implement with native browser apis, asking the robots

Done, great idea, thx! https://github.com/franciscocarloserra/p2p-call/commit/c722d...

Using the vision feature feels so token bloated to me. I generally delegate that to a fresh context if possible.

Models + harnesses are simply becoming too useful, it's a skill issue. Especially as new models incorporate knowledge about the harness and workflows on their training data/internal representations.

With the rise of better and better edge inference and almost state-of-the-art open-weights models, I don't see this going well. Almost any additional latency makes talk-back less usable, so it's reasonable for it to move out of the cloud. And TTS models are becoming trivial to replicate for informational uses.

Super cool dataset. Thank you very much. You just saved Hacker News, some bandwidth and server compute as well. Cool to see the trends go up and down as well. Do you have ideas on any additional dimensions it would be easy to explore as well?

This is super interesting and potentially extremely useful as well. What is the performance improvements against doing it via API remoting or translating?

I'd be super interested in knowing how other models perform by default and knowing the exact details of the prompts as well and seeing how two agents with different contexts or isolated contexts perform against one another, one as an implementer and another one as a validator.

This actually has benchmarking potential. It's especially interesting now that some of the models start having knowledge about themselves in the training dataset.

is there any published code or prompts regarding this for the either the prompts or the evals?

*also trying it against several different delegation topologies would be super interesting but it seems like none of the harnesses by default offer control over how many levels of delegations and which parts of the task each instance or each agent is gonna focus on